Running Out of Room for Models — I Checked SSD Prices and Found Something I Should Have Looked at First

September 13, 2026September 9, 2026

6 August 2026.

My free disk space had dropped to 211GB.

Out of 1.5TB, I’d used up 1.2TB. Writing local-AI articles for this blog means models keep piling up: you download one to measure it, keep it around to compare against later, and then a new one shows up and you download that too. Deleting them would solve it, but since I sometimes go back and re-measure, I can never quite bring myself to.

Figuring it was time to add more storage, I went to check prices. What I found wasn’t what I expected.

While I was at it, I also measured what was actually going on with my own machine, and that turned out to be the more surprising result. Isn’t there something worth checking on your own hardware before buying more storage? Does buying a faster SSD actually shorten the wait?

What’s Actually Taking Up So Much Space?

First, I counted up what was actually in there.

Here’s the size of individual models too, for the larger ones I have on hand.

Keep three of the large ones around and that’s 200GB right there. In my case, I keep a range of sizes on hand for side-by-side comparisons, which is how it piled up like this.

What Do SSDs Actually Cost Right Now?

I went in planning to buy more, and stopped short.

That’s close to ¥40,000 for 2TB — the cheap end of the range, as checked on Amazon on 6 August 2026. From a well-known brand, 2TB runs from the ¥60,000s into the ¥80,000s.

Looking into why led to the same place as the GPU price hikes.

“Wait for it to get cheaper" isn’t a great strategy right now. That said, buying something expensive doesn’t necessarily solve the problem either, as it turns out.

Does Buying a Faster SSD Actually Shorten the Wait?

From here, this is what I measured myself.

Store listings show numbers like “up to 7,400MB/s read." That means it can read 7.4GB per second. Buy one with a bigger number, and model loading looks like it should speed up.

So I measured how many GB/s the SSD I actually use can read, bypassing the cache (O_DIRECT), under three different conditions.

It came out to about 1GB/s. That doesn’t change no matter how many files you read at once.

That’s a seventh of the 7,400MB/s figure on the box. Granted, what I have is a few years old, and a newer drive would deliver more. But on this machine, it’s the 1GB/s figure that’s setting the wait time, and plugging in a 7,400MB/s drive doesn’t guarantee you’ll actually get that number. The reason is in the next section.

How Many PCIe Lanes Is My SSD Actually Connected With?

Digging into this turned up something I didn’t expect.

An M.2 SSD connects to the computer through a pathway called PCIe. That pathway has lanes, and the usual count is four (x4). But when I checked my own SSD, it was running on two lanes (x2). The drive itself supports four lanes, and only half were actually being used.

A motherboard’s M.2 slots can have different lane counts depending on position, and some share lanes with SATA ports. It’s written in the manual, but it’s not something you tend to pay attention to when building a machine.

The biggest takeaway from this: before buying a faster SSD, it’s quicker to check how many lanes the one you already have is connected with. Moving it to a different slot costs nothing.

On Linux, you can check with the following commands.

cat /sys/class/nvme/nvme0/device/current_link_width
cat /sys/class/nvme/nvme0/device/max_link_width

The first shows how many lanes are in use now; the second shows the maximum it can go up to. If the two numbers differ, you’re losing lanes.

TLC vs QLC: What Actually Changes?

This is about what’s inside the SSD. It turned out to connect directly to both price and lifespan.

An SSD stores data by holding a charge in small cells. What the cell is called depends on how many bits get packed into each one.

The more you pack in, the cheaper and larger the capacity gets, and the shorter the lifespan gets in exchange. QLC, which has to distinguish between 16 voltage levels, needs precise control on both writes and reads, and that burden accumulates over time.

The important point here is that reading and writing are completely different stories.

Checking through the best-selling 2TB drives, almost all of them turned out to be TLC. QLC has fallen out of the mainstream lineup — possibly because rising prices have narrowed the gap, eroding cheapness as QLC’s one selling point.

For what it’s worth, both of my drives are QLC. That apparently puts me in the minority these days. Even so, model loading has been working just fine.

Lining Up the Best-Selling 2TB Drives

Checked on Amazon on 6 August 2026. TBW is a rough measure of “how many TB you can write in total."

TBW lined up in reverse of price. The cheapest, Ediloca, is rated at 1,400TB; the ¥68,000 WD comes in at 1,200TB. It doesn’t line up as “the pricier one lasts longer."

There was also a case of conflicting figures. Silicon Power’s product page lists “TBW: 1200" in the description, but says “warranty within a TBW value of (300TB)" in the warranty terms — a 4x difference on the same page. My take is that the number in the warranty terms is the more reliable one to go by.

Of the drives in the table, WD Black SN7100 was the one with a clearly stated TBW figure and a price I could confirm.

As an Amazon Associate we earn from qualifying purchases.

What Actually Wears Out a Drive’s Lifespan?

TBW is “total writes." This turned out to be the most reassuring part of all this.

Reading barely touches lifespan. Load a model as many times as you like, and it doesn’t eat into that number.

Storage for local-AI models is a “write once, read many times" kind of use. Even downloading 100 models at 40GB each only comes to 4TB written — still just 0.3% of a 1,400TB rating.

I also checked how much my own SSD had written over the 16 days since it last booted.

For this kind of use, I couldn’t find a reason to pay more out of concern for lifespan.

Which Habits Burn Through Lifespan Faster?

• Using it as the overflow destination for what doesn’t fit in memory — small writes keep happening endlessly
• Generating intermediate files for things like additional training (LoRA)
• Repeatedly downloading and deleting models. At 40GB each, that adds up every time you swap one out
• A resident tool that keeps writing logs continuously

The last one has a real-world example where it actually became a problem.

A bug was found in a command-line tool for AI-assisted coding, where it kept writing logs at the most verbose setting. In the reporter’s environment, that came to 37TB written over 21 days. Converted to a year, that’s about 640TB — enough to use up a 1TB-class SSD’s lifespan (around 600TB) in under a year.

I checked my own setup too, and the relevant log was 48KB — this bug wasn’t happening here. That’s also consistent with the roughly 7TB-a-year write figure above.

“Doing local AI makes me worry about SSD lifespan" is better aimed at the tools running constantly in the background than at storing models. On Linux, you can check your own write volume with this.

cat /sys/block/nvme0n1/stat | awk '{print $7*512/1024/1024/1024, "GB"}'

That’s the cumulative total since boot. Divide by how many days it’s been running, and you get a rough per-year estimate.

At What Point Does Capacity Actually Run Out?

I’m using 742GB, but that’s because I keep a range of sizes on hand for side-by-side comparisons — it’s not typical usage.

I worked out how many models fit in 1TB, calculating on the basis that about 900GB is actually usable.

Even 1TB is enough to get started. Up to the 30B class, you can fit dozens of models.

There’s also this: the size worth keeping around is decided by the memory you actually have. Putting an 86GB model on a machine with 12GB or 24GB of VRAM doesn’t get you anything usable. Since there’s no need to keep what you can’t load anyway, the capacity you actually need stays naturally in check.

You start running short once you want to keep several 70B-and-up models around. Even then, there’s an option to split storage across locations before jumping straight to buying more.

I haven’t measured the external drive yet, so the number above is the spec sheet figure. That’s next time’s homework.

Here are external SSDs supporting USB 10Gbps (USB 3.2 Gen2). Checked for stock on Amazon on 5 September 2026.

How Does the Price per GB Change With Capacity?

Price per GB showed a clear difference by capacity: ¥20.0/GB at 1TB, ¥18.5/GB at 2TB, ¥16.8/GB at 4TB.

Buying bigger is undeniably the better deal per gigabyte. That said, 1TB is about ¥20,000 and 4TB is about ¥67,000, so whether you can pay three times as much right now is a separate question. Adding capacity later also means the hassle of swapping drives, so take this as no more than “go bigger if your budget allows."

Here’s a 4TB NVMe SSD, checked on 5 September 2026. Prices kept climbing, and it had gone up even further from the roughly ¥67,000 seen earlier in this article as of 6 August.

What I Did Not Measure

Prices were checked on 6 August 2026. Given that prices keep climbing, they may well have changed by the time you’re reading this. Speed is a value measured on a single drive of mine, not an average across repeated measurements. I haven’t measured other SSDs or HDDs, so no comparison there.

In Summary: So, Which One Should You Actually Buy?

Before looking into this, I assumed “pick the faster one and loading gets faster." Once I measured it, that wasn’t really the case.

• Check how many lanes what’s currently installed is using before you buy. Mine was running on half. Just moving it to a different slot costs nothing
• TLC is enough. Nearly all the best-selling 2TB drives are TLC anyway, so there isn’t much room to choose otherwise
• TBW isn’t proportional to price. The cheaper option sometimes has the bigger number, and figures can differ depending on where they’re written (description vs. warranty terms)
• Reading doesn’t wear down lifespan. Used as storage, that’s not going to be a problem
• Capacity is decided by the memory you actually have. There’s no point keeping a model you can’t load anyway
• You can start with 1TB. Once it runs short, there’s the option of offloading to external drives or an HDD
• Bigger is cheaper per gigabyte. Whether you can pay three times as much right now is a separate call

As for what I’m actually going to do: first, move the drive to a different slot and see if the lane count goes up. If that’s not enough, I’ll buy more, but the free option comes first.

Given prices are climbing right now, not having to buy at all is the best outcome.

Products Covered Here

If you just want one drive: a 2TB unit for storage with a clearly stated TBW figure.

The 4TB that comes out cheapest per gigabyte.

For when the internal slots are full: external options.

How I Measured, and an Honest Caveat

Read speed was measured bypassing the cache (O_DIRECT). I checked it four ways: one file at a time, 4 at once, 8 at once, and reading 39.6GB straight through after flushing the cache. All four came out at about 1GB/s.

I did not measure the write slowdown (the drop in speed once the SLC cache is used up). Actually measuring that would take close to 100GB of writes, which means spending lifespan to measure lifespan. Those figures come from each maker’s own explanations.

Prices, ratings, and TBW were checked on Amazon’s product pages on 6 August 2026. Since prices keep climbing, they may have changed by the time you’re reading this. For the NAND trading price and the price hikes from various makers, I quoted reported figures with their sources noted — I did not measure these myself.

*Test environment: desktop PC (GeForce RTX 3090 24GB + GeForce RTX 3060 12GB / 62GB system memory / two NVMe SSDs, two HDDs). As of August 2026.

Here’s an article that sorted storage options by price and speed.

My SSD Filled Up — Rethinking Where Data Goes, by Price and by Speed

18 August 2026 Data kept piling up, and the SSD inside my PC is filling up. Local AI models alon ...

 https://kimono-oyaji.com/storage-hdd-nas-options-en/

If you want to pick again starting from a full GPU list, here’s that one.

2026 GPU Spec List for Local AI, VR and Image Generation

How to choose a GPU and why VRAM matters so much are covered in separate articles. Related For h ...

 https://kimono-oyaji.com/gpu-spec-list-2026/